ARM: Do not set FPSCR when converting to clamped uint8

Setting the FPSCR flags is expensive on some CPUs. Get rid of repeated
setting of the FPSCR by relying on the correct default flags being set
when doing uint8 clamping. Also use vcvt_u32_f64 instead of vcvt_s32_f64,
which enables removing the check against zero (vcvt_u32_f64 will clamp to
zero).

To be on the safe side, add asserts to check that the VFP rounding mode
flags are set to default as expected.

This increases performance of a hot loop repeatedly setting
Uint8ClampedArray values on some CPUs by as much as a factor of 12.

BUG=v8:3253
LOG=N
R=jacob.bramley@arm.com, rmcilroy@chromium.org, ulan@chromium.org

Review URL: https://codereview.chromium.org/230473005

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@20676 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/arm/macro-assembler-arm.cc

@@ -796,6 +796,10 @@ void MacroAssembler::VFPEnsureFPSCRState(Register scratch) {
   // If needed, restore wanted bits of FPSCR.
   Label fpscr_done;
   vmrs(scratch);
+  if (emit_debug_code()) {
+    tst(scratch, Operand(kVFPRoundingModeMask));
+    Assert(eq, kDefaultRoundingModeNotSet);
+  }
   tst(scratch, Operand(kVFPDefaultNaNModeControlBit));
   b(ne, &fpscr_done);
   orr(scratch, scratch, Operand(kVFPDefaultNaNModeControlBit));
@@ -3800,36 +3804,19 @@ void MacroAssembler::ClampUint8(Register output_reg, Register input_reg) {
 void MacroAssembler::ClampDoubleToUint8(Register result_reg,
                                         DwVfpRegister input_reg,
                                         LowDwVfpRegister double_scratch) {
-  Label above_zero;
   Label done;
-  Label in_bounds;
-
-  VFPCompareAndSetFlags(input_reg, 0.0);
-  b(gt, &above_zero);
-
-  // Double value is less than zero, NaN or Inf, return 0.
-  mov(result_reg, Operand::Zero());
-  b(al, &done);
 
-  // Double value is >= 255, return 255.
-  bind(&above_zero);
+  // Handle inputs >= 255 (including +infinity).
   Vmov(double_scratch, 255.0, result_reg);
-  VFPCompareAndSetFlags(input_reg, double_scratch);
-  b(le, &in_bounds);
   mov(result_reg, Operand(255));
-  b(al, &done);
-
-  // In 0-255 range, round and truncate.
-  bind(&in_bounds);
-  // Save FPSCR.
-  vmrs(ip);
-  // Set rounding mode to round to the nearest integer by clearing bits[23:22].
-  bic(result_reg, ip, Operand(kVFPRoundingModeMask));
-  vmsr(result_reg);
-  vcvt_s32_f64(double_scratch.low(), input_reg, kFPSCRRounding);
+  VFPCompareAndSetFlags(input_reg, double_scratch);
+  b(ge, &done);
+
+  // For inputs < 255 (including negative) vcvt_u32_f64 with round-to-nearest
+  // rounding mode will provide the correct result.
+  vcvt_u32_f64(double_scratch.low(), input_reg, kFPSCRRounding);
   vmov(result_reg, double_scratch.low());
-  // Restore FPSCR.
-  vmsr(ip);
+
   bind(&done);
 }
 

src/objects.h

@@ -1127,6 +1127,7 @@ class MaybeObject BASE_EMBEDDED {
   V(kDeclarationInCatchContext, "Declaration in catch context")               \
   V(kDeclarationInWithContext, "Declaration in with context")                 \
   V(kDefaultNaNModeNotSet, "Default NaN mode not set")                        \
+  V(kDefaultRoundingModeNotSet, "Default rounding mode not set")              \
   V(kDeleteWithGlobalVariable, "Delete with global variable")                 \
   V(kDeleteWithNonGlobalVariable, "Delete with non-global variable")          \
   V(kDestinationOfCopyNotAligned, "Destination of copy not aligned")          \